Geodeziya, kartografiya, geografiya

Which classification method should I use?
The classification process involves translating the pixel values in a satellite image into meaningful categories. In the case of land cover classification these categories comprise different types of land cover defined by the classification scheme that is being implemented. There are dozens, if not hundreds, of classification methods that can be used to group image pixels into meaningful categories. Unfortunately there is not a single "best" approach to image classification. The choice you make depends a lot on the algorithms that are available to you with the image processing software you use and your familiarity and experience with the different methods. To facilitate the discussion of the different methods we will group them into automated, manual, and hybrid approaches.
Automated
The majority of classification methods fall in this category. With this approach an algorithm is used to assign individual pixels or groups of pixels to one of the valid categories. One advantage of automated approaches is that the algorithm is applied systematically throughout the entire image relatively quickly. The algorithm can also utilize many more layers than a human using visual methods. For example, using manual methods an interpreter will often be limited to viewing a 3-band image whereas using automated methods it is possible to incorporate data from hundreds of satellite bands. Traditionally, automated algorithms tended to be limited to using only the pixel values in the image but there are now several approaches that allow a user to easily incorporate ancillary data layers such as elevation, slope, aspect, soil type, and a host of other biophysical layers to improve the classification accuracy.
In order for an automated classification algorithm to associate pixel values with the correct land cover category, some input from an analyst is necessary. When this information is provided before the algorithm is run the procedure is referred to as supervised classification. With this approach the user identifies sample pixels in the image that can be used as representative examples for a particular land cover category and then the sample pixels are used to train the algorithm to locate similar pixels in the image. When a supervised classification is run the result is a land cover map with all of the pixels labeled as a particular cover type. The other way to classify an image is to start by letting the computer group similar pixels together into unlabeled classes (clusters) and then have the analyst label the clusters with the appropriate land cover category. This approach is called unsupervised classification since the algorithm works without a-priori input of information about existing land cover samples.
Both supervised and unsupervised classification methods can produce reliable results; however, there is greater variety in the available algorithms geared toward supervised classification. The primary difference between classification algorithms is the way in which they determine how an individual pixel is assigned to a land cover category.
We present a short overview of some of the more popular classification approaches to provide you with insight into the range of possibilities. Many of these algorithms come from the field of machine learning and they can be quite complex. Providing details of individual algorithms is beyond the scope of this guide but the basic concept is explained in the Feature Space Interactive. More detailed information about classification algorithms is available in remote sensing textbooks, training courses, or on the Internet. The best way to get a feel for how these different algorithms work is to practice applying them using remote sensing software packages. Although it is not required, an understanding of how an algorithm works will help the analyst use the selected method more effectively. Again, talking with someone with experience with the different classification approaches can be helpful in deciding which one to use.